Nerve is a simple but powerful Agent Development Kit (ADK), designed to help users build, run, evaluate, and orchestrate agents based on large language models (LLMs). It is primarily aimed at technical users and emphasizes programmable, auditable, and reproducible automation.
Main capabilities include:
- Declarative agents: Define agents using a clear YAML format, with system prompts, tasks, tools, and variables all centralized in one file.
- Built-in tools and extensibility: Use shell commands, Python functions, or remote tools to give agents capabilities. Tools are fully typed and annotated.
- Native MCP support (client and server): Nerve is the first framework that allows defining MCP servers in YAML, and can run as both a client and a server, thereby supporting agent teams and deep orchestration.
- Evaluation mode: Benchmark agents using YAML, Parquet, or folder-based test cases. Run reproducible tests, record structured outputs, and track regressions or progress.
- Compose agents into simple linear pipelines to create multi-step automations with shared context.
- Built on LiteLLM, supporting OpenAI, Anthropic, Ollama, and dozens of other models, switchable with just one line.
Usage: Define agents via YAML, and use the CLI to run, evaluate, and orchestrate. Official documentation and examples are provided for further understanding.
Applicable scenarios: Suitable for technical users who need to use LLMs to build auditable, reproducible automation workflows, such as multi-step task automation, agent orchestration, and benchmarking.
Limitations: Specific pricing, fees, or platform restriction information is not currently provided. The project is released under the GPL 3 license.
Other information: The project welcomes contributions, and has contribution guidelines and a Discord community.
